This talk reviews new directions for cyber security built around machine learning and using adversarial learning and conformal prediction to enhance network and computing services defenses against adaptive, malicious, persistent, and tactical threats. The motivation for using conformal prediction and its immediate offspring, those of semi-supervised learning and transduction, comes from them supporting discriminative and non-parametric methods using likelihood ratios; demarcation using cohorts, local estimation, and non-conformity measures; randomness for hypothesis testing and inference using sensitivity and stability analysis; reliability indices on prediction outcomes using credibility and confidence to assist meta-reasoning and information fusion; and open set recognition including novelty detection and the reject option for negative selection. The solutions proffered are built around active learning, meta-reasoning, randomness, semantics and stratification using topics and most important and above all using adaptive Oracles that are effective and valid for the purpose of model selection and prediction. Effective to be resilient to malicious attacks aimed at subverting promptness, selective in separating the wheat (e.g., informative patterns) from the chaff (e.g., obfuscation), and valid and well - calibrated to not overrate the accuracy and reliability of the predictions made.
展开▼
